Artificial Intelligence Techniques for Observation of Earth’s Changes

Abstract

This chapter discusses the primary components that contribute to the observation of Earth’s changes, including Land Observation Satellites, land classification techniques and their stages of development, and Machine Learning Techniques. It will give a comprehensive summary of the development stages of high-resolution satellites. It also details land classification with artificial intelligence algorithms. It will also give knowledge of classification methodologies from various Fundamentals of Machine Learning Classifiers: Pixel-based (PB), Sub-pixel-based (SPB), Object-based (OB), Knowledge-based (KB), Rule-based (RB), Distance-based (DB), Neural-based (NB), Parameter Based (PB), object-based image analysis (OBIA). It includes several different classifiers for LULC Classification. This chapter will include two applications for land observation satellites: The first is land use and land cover change observation with a practical example (study land use and land cover classification for Sana’a of Yemen as a case study from 1980 to 2020). The second application is satellite altimetry monitoring changes in mean sea level. The most significant contributions of it are the integration of these components. This chapter will be crucial in helping future researchers comprehend this topic. It will aid them in selecting the most appropriate and effective satellites for monitoring Earth’s changes and the most efficient classifier for their research.

1. Introduction

The art and science of measuring the planet earth through sensors or satellites are known as remote sensing, which, together with GIS technology, become an essential aid in collecting data about the Earth. The overall purpose of image collection is to naturally classify all pixels in an image into land cover groupings or subjects. In LULC categorization, their unique Artificial Intelligence approaches play a key role. Therefore, the work is interesting for this book. The work represents an overview of the application of artificial intelligence in detecting Land Use Land Cover (LULC) changes on Earth. This chapter reviews Earth observation satellites and their development, brings synthetic intelligence procedures for Land Use Land Cover (supervised and unsupervised methods) and fundamentals of ML classifiers, and ends with the challenges of AI techniques for LULC classification and conclusion [1].

The land has become a gigantic and immense resource of economic that cannot be underestimated in any region, where Earth Changes Observation led to serving the country’s economic, political, and social needs. Understanding land changes is vital for land resource management and assessing the technology’s potential, where LULC change detection assists policymakers in understanding the dynamics of environmental change to ensure long-term growth. As a result, LULC feature identification has become an essential research topic, necessitating the development of a robust and reliable LULC classification [2].

Land use cover is necessary to make up the land’s physically present and visible surface components [3] which allow researchers to investigate landscape patterns and features, which are essential to understanding land size, location, and condition of the size, structure, and state of the ecosystem [4]. The importance of land classification stems from using a specific piece of land that may be linked to significant price differences, necessitating a well-defined land categorization. The price (development) of land underneath houses, for example, may be drastically different from the price (impact) of agricultural land [5]. As a result, remote sensing and geographic information systems (GIS) have become essential aid in collecting data about the Earth, and It’s considered critical for risk assessment and monitoring land degradation and conservation [6, 7].

It is possible to collect data across large geographical areas and define natural qualities or physical items on the ground. Analyzed surface areas and objects regularly, tracked their changes over time and combined this data [8] by several decision support systems relying on remote sensing for land use and land cover (LULC) detection [9]. As considered, Artificial intelligence (AI) is a technique solid and active in studying and developing computers or computational systems that can accomplish tasks that would need human intelligence in this field. According to the innovative operation, artificial intelligence algorithms play an essential part in LULC, where these classifiers for LULC classification can be split into different classes. Arrangement strategies are divided into two categories. The first category is traditional machine learning (complex classification) (unsupervised and semi-supervised, supervised), contemporary machine learning, and based knowledge discovery are examples of challenging classes. The second category is soft classification [10].

AI is useful because it enables software to execute human-like functions like reasoning, planning, communication, and perception more effectively, efficiently, and at a lower cost [11]. Quantum computing has much potential for improving AI and machine learning algorithms. Although the technology is currently out of reach for most people, Microsoft, Amazon, and IBM are making quantum computing resources and simulations available via cloud models [12].

2. Satellites for land observation

Earth changes observation is a ground photography scanning system that collects, stores, analyses, and displays land photographs using remote sensors at regular intervals [13]. Earth changes observation is used to detect changes in land cover over time and monitor and analyses changes in the natural and built environments through the land’s physical, computational, and biological systems, among other things, according to the Land Observation Group [7].

Land observation satellites are satellites designed to view the Earth from orbit and are used for various reasons, including mapping, environmental monitoring, meteorology, and other applications. They typically include remote sensors and wireless devices. Ground imaging satellites, which capture photos from satellites, are the most frequent type [8].

The spread of satellite launches in most technologically advanced countries has led to a new shift, in general, exploring land uses, land cover, and earth changes observation [14]. So the increasing need to continue developing remote sensing satellites to monitor the ground and know the current limitations to launch a revolution in space and technological development requires a careful study of the capabilities and challenges. This study will provide a comprehensive survey in this field to find valuable details [15].

2.1 Types of land observation satellites

The resolution of an image refers to the potential detail provided by the imagery. Resolution refers to the smallest size an object or element can be represented in an image. Higher resolution means that pixel sizes are smaller, which provides more details. Then Figures 1 and 2 show the types of resolution of Land Observation Satellites (LOS) with some significant features [15].

Figure 1.

Types satellites of land changes detection.

Figure 2.

Types of resolution satellites.

2.2 Development of land observation satellites (LOS)

The spatial resolution was the fundamental distinction between high-resolution and low- and medium-resolution satellites. The higher spatial resolution (0.5–1.5 m/pxl) elevates the image qualities - from an unexpectedly detailed image due to the length units to a ratio of pixel values that provides the user with greater precision. It is also surrounded by the latest Optical technologies onboard the high-resolution satellites for remote sensing in addition to a high visit (≤ one day), which allows for observing the current conditions of the Earth’s surface [16]. Another great feature of high-definition satellites for commercial data is the assignment of a dedicated mission in terms of the ability to commission a high-resolution satellite to take a new photo. Always the reason why HRSI has a unique characteristic is that it is clearly and publicly commercially based data (here applied to the saying: you get for what you pay [17]. The fact that high-resolution satellites are commercial offers them the following advantages:

• Filtering the image to 100% of the region covering your use case.

• Get free samples, and a sneak peek.

• Instant access to the pricing because it is determined automatically.

• Three working days to receive the image.

• When used offline, high-resolution photos can be processed directly in the browser, avoiding FTP and downloading high-resolution satellite images for further analysis. It saves time, effort, storage space, and the cost of specialized tools.

The literature describes the benefits of high-resolution satellites, quick delivery, and fine details [18]. After the appearance of high-resolution satellites was removed, this section presents an overview of the high-resolution expected. The following are details on the development of (LOS) in Tables 1 and 2: (1) from 1999 to 2010, the first generation. (2) 2010–2015, the second generation. (3) 2015–2020 third generation (4) The Future—Fourth Generation 2020 As stated in the references [19], the data was gathered from a variety of credible and recent sources. The investigations indicated that new and exciting breakthroughs would arise. Many satellites with high spatial accuracy will occur in the next century’s early years. The inaccuracy of high-resolution spatial image data is better, but it is also more expensive [20]. However, there is still a demand for medium-resolution satellites because of the benefits that commercial satellites have and lack [21]. They’re also less priced and arrive faster. It was also mentioned that the launch of more remote sensing satellites with higher accuracy than currently available will have increased. Tables 1 and 2 [22] provide details on high-resolution satellite launches.

Country	Spacecraft	Date	Resolution
USA	IKONOS	1999	1 m
	QuickBird-1	2000	0.8 m
	EROS-A1	2000	1.5 m
	Preview-3	2001	1 m
	Preview-4	2001	1 m
	EROS-A	2001	1 m
	Quick Bird	2001	0.61 m
	WorldView-1	2007	0.5 m
	WorldView-2	2009	0.5 m
	GeoEye-1	2008	0.46 m
INDIA	IRS-P5	2002	2.5 m
	IRS-P6	2001
	IRS-2A	2003	1 m
	Cartosat-2	2007	0.8 m
	Cartosat-2A	2008
	Cartosat-1	2005	2.5 m
Russia	Kometa-20	2000	2 m
Canada	Radarsat2	2002	3 m radar
China/Brasil	CBERS-3	2002	3 m
China/Brasil	CBERS-4	2002	3 m
Japan	ALOS	2002	2.5 m
France	SPOT-5	2002	5 m
Taiwan, land & ocean	ROCCAT-2	2002
DLR, Radar	TerraSAR,	2004	1 m
Emirates	DubaiSat-1	2009
Italy	CSK-1	2007
Italy	CSK-2	2007
Italy	CSK-3	2008
South Korea	Kompakt-2	2006
Country	Spacecraft	date	Resolution
Canada	RADARSAT-2	2007
Spain	(SEOSat)	2007
USA	WorldView-3	2014	3 0.4 m PAN
	SkySat —1	2013	0.8 m PAN and 1.0 m MS
	SkySat —2	2014	0.8 m PAN and 1.0 m MS
	Alat-2A	2010	2.5 m PAN and a 10 m MS
	Sentinel-1	2014
	VNREDSat-1A	2013	2.5 m
CHINA	Gaofen,GF1	2013
	Gaofen, GF2	2014	0.8 m and a MS 3.2 m.
	China-Brazil-Earth Resources-Satellite (CBERS)-4	2014
	TripleSat-1 to 3	2015
INDIA	Cartosat-2B	2010
Algeria	AlSat-2A	2010
Italy	CSK-4	2010
Spain	Deimos-2	2014
Emirates	Dubai’s-2	2013
Kazakhstan	(KazEOSat-1)	2014
South Korea	Kompakt-3	2012
South Korea	Kompakt-3A	2015
Nigeria	Nigeria-2,	2011
France	Pleiades-1A	2011
France	Pleiades-1B	2012
France, Azerbaijan	SPOT-6	2012
France, Azerbaijan	SPOT-7	2014
United Kingdom, China	TripleSat-1, —2,-3	2015
Vietnam	VNREDSat-1A	2013	2.5 m PAN and 10 m MS
USA	WorldView-4	2016	31 cm PAN and 1.24 m MS
	SkySat —3	2016	0.8 m PAN and 1.0 m MS
	SkySat —4 to —7	2016	0.8 m PAN and 1.0 m MS
	SkySat −8 to −13	2017	0.8 m PAN and 1.0 m MS
EUROPEAN	Alat-2B	2016	2.5 m PAN and a 10 m MS
	Sentinel-1B	2016
	Sentinel-2A	2015
	Sentinel-2B	2017
	Sentinel-3A	2016
	Sentinel-3B	2018
CHINA	Gaofen, GF4	2015
	Gaofen, GF3	2016
	Gaofen,GF5,GF6	2017
	Gaofen, GF7	2018
	Zhuhai	2018
	SSTL-S1–4	2018	1 m PAN mode and m MS
INDIA	Cartosat-2C	2016
	Cartosat-2D	2017
	Cartosat-2E	2017
	Cartosat-2F	2018
	Cartosat-3	2019 (planned)
	GEO Imaging Satellite (GISAT)	2019	42 to 318 m
	HRSAT 1A,; 1B, and 1C	2020 (planned)
	resources-3S	2019 (planned)
	resources-3SA	2020 (planned)
Japan	ALOS-3	2020 (planned)
Algeria	Alat-2B	2016
South Korea	CAS500–1 and — 2	2020 (planned)
Italy	CSG-1	2019 (planned)
Italy	CSG-2	2020 (planned)
United Arab Emirates	Khalifasat	2018
Morocco	Mohammed VI-A	2017
Morocco	Mohammed VI-B	2018
Argentina	The Aleph-1	2018
Peru	Peru’s-1	2016
Canada	RADARSAT,	2019
Kingdom, China	SSTL-S1	2018
United Kingdom	Vivid-i 1 to 5	2019
Venezuela	VRSS-2	2017

Table 1.

High-resolution optical space sensors for the first generation, 1999–2020.

MS = multispectral and PAN = panchromatic [2].

Country	The Spacecraft	Date
USA	(1HOPSat)	2020 (planned)
EUROPEAN	(FLEX)	2022 (planned)
	Sentinel-1C	2021 (Planned)
	Sentinel-1D	2023 (Planned)
	Sentinel-2C	2020 (planned)
	Sentinel-2D	2021 (planned)
	Sentinel-3C	2020 (planned)
	Sentinel-3D	2022 (planned)
INDIA	NISAR	2021 (planned)
South Korea	Kompakt-7	2021 (planned)

Table 2.

High-resolution optical space sensors for the fourth generation, 2020-future [2].

3. AI approaches to analyze LULC

Recently, Classifiers that create exact LULC maps have been in high demand. Dependable Information is required from remotely sensed pictures, even on high-dimensional, complex data. Machine Learning Classifiers have a significant role in giving good classification results. Several aspects influence the accuracy of classified maps, including training sample size, training sample quality, thematic correctness, classifier choice, study region size, etc. Understanding these criteria will aid in achieving the highest classification accuracy feasible for a given need [23]. Big Data challenges arise when classification tasks involving multiple satellite photos and features become computationally intensive. To identify AI techniques for LULC of ML, as described in Figure 3, this section from this chapter searched deeply for several methods that supply a significant impetus for future readers to develop ML techniques. As stated in Table 3, we investigated the foundations of different AI classification algorithms in this chapter.

Figure 3.

Types of AI techniques for LULC classification.

To provide new readers knowledge about different LULC system foundations as follows: Pixel-based (PB), Sub-pixel-based (SPB), Object-based(OB), Knowledge-based(KB), Rule-based(RB), Distance-based(DB), Neural-based(NB), Parameter Based(PB), object-based image analysis (OBIA). It includes several different classifiers for LULC that are as follows: (Random Forst -RF, Tree Decision Classification -TDC, Maximum Likelihood Classifier –MLC, Spectral Angle Mapper Classification -SAM, Support Vector Machine -SVM, K-Nearest Neighbor – KNN, Minimum Distance Classification - MDC, Artificial Neural Networks - ANN, Mahalanobis, Maximum Entropy, Parallelepiped, Boosting, Normal Bayes, ISOData, and K-means) [24].

3.1 Supervised methods

Classification Supervised (human-guided): This is based on the idea that a user can select sample pixels in an image representing different classes and then tell image processing software to use these training sites as references when classifying the rest of the pixels in the image [25]. The user’s knowledge is used to choose training locations (testing sets or input classes). The user defines the boundaries for how similar they must be to group pixels together. These bounds are usually determined using the spectral properties of the training zone, plus or minus a defined increment (sometimes based on “brightness” or strength of reflection in specific spectral bands). In supervised learning [26], you use well-labeled data to train the algorithm.

It signifies that certain information has already been tagged with the appropriate response. It’s analogous to learning with a teacher or supervisor present. A supervised learning system that learns from labeled training data can predict unexpected data outputs. It is possible to design, scale, and deploy accurate supervised machine learning. A group of highly skilled data scientists must devote time and technical skills to build a data science model. Data scientists must also keep their models up to date to ensure accurate insights even if the data changes [27]. The different supervised approaches are shown in Figure 4 [27]: classification and regression.

Figure 4.

Classification and regression from supervised techniques.

3.1.1 Type of supervised methods

Regression: A method for predicting a single output value using training data, for example, uses regression to predict a property’s price based on training data. Other input variables include location, dwelling size, and other aspects [28].

Classification: “classification” refers to categorizing output into different groupings. Binary type is when an algorithm splits data into two groups. The multiclass category is choosing between more than two classifications [29].

3.2 Unsupervised methods

The model is not supervised in unsupervised learning. Instead, allowing the model to determine what it requires would be beneficial. It mainly deals with data that has not been labeled. Unsupervised learning algorithms allow for more complex processing tasks than supervised learning algorithms. On the other hand, unsupervised learning is potentially more unpredictable than natural learning systems such as deep learning and reinforcement learning [30]. Classification Unsupervised (according to software): This classification is based on software analysis of an image without using user-supplied example classes. The computer uses algorithms to determine which pixels are connected and categorize them accordingly. Figure 5 [31] shows this well.

Figure 5.

Types of unsupervised techniques.

3.2.1 Type of unsupervised method

Clustering When it comes to unsupervised learning, clustering is a crucial idea. Its primary purpose is discovering a structure or pattern in uncategorized data [31]. If natural clusters (groups) exist in the data, clustering algorithms will analyze and find them [32].

Associated: Associated: You can create associations between data elements in massive databases using association rules. This unsupervised method searches extensive databases for intriguing correlations between variables. For example, people who purchase a new home are more likely to buy new furnishings [33].

3.2.2 Comparison between supervised and unsupervised methods

There are several different Supervised and Unsupervised Methods methods, explained in Table 3 down [34]:

3.3 The fundamentals of ML classifiers

3.3.1 Pixel-based classification

Pixel depiction approaches, such as the model remote recognizing image request technique, assume that each Pixel is pure and is commonly referred to as a single land spread type. Using this technique, distantly identifying imagery is perceived as a collection of pixels containing alarming data. As a result, extra standard components and their changes (for example, head sections, vegetation records, and so on) represent a pixel classifier commitment. Pixel-wise portrayal estimations can be divided into autonomous and supervised game plans. Using single classifiers, far distinct image is sorted into multiple classes based on the trademark groupings of the image without the usage of ready data or primary data on the review zone [35].

3.3.2 Sub-pixel-wise based

According to pixel-wise, far from identifying picture-gathering processes, each image pixel has only one land use kind. Regardless, such a notion is typically erroneous for medium and coarse objective imaging due to the unpredictability of scenes when viewed about the spatial target of a distant identifying picture. On the other side, using hard-plan plans decreases land use spread and helps representation accuracy. Because the areal degree of each land use may be precisely quantified, subpixel collection technologies are a popular alternative. Primary subpixel depiction backslides displaying backslide tree analysis, and supernatural mix assessment have been designed to address the mixing pixel issue. Overall, each Pixel receives fragmented enrolments with the soft depiction, and the contrasting areal degree of each class may be assessed [36].

3.3.3 Object-based classification

The spatial features of each Pixel as they relate to one another are considered while classifying a small collection of pixels. A pixel collection would be used as a preparation model for the classification algorithm. The classification algorithm would produce a class forecast for pixels. Object-based approaches divide images and route image requests to things rather than pixels, resulting in picture conflicts. In picture division approaches, mysterious, spatial, textural, and crucial information highlight picture objects. These articles have also mastered using unnatural and other critical models. Object-based techniques have greatly improved accuracy in multiple investigations [37] because different picture pixels make up a geographic item.

3.3.4 Image segmentation and object-based image analysis (OBIA)

A high-level image is broken down into numerous homogeneous components, each distinct. Image and article-based picture divisions, utilized in content-based image recovery, clinical imaging, object revelation, and other areas, are the third and fourth critical social events. Furthermore, Kitting and Land grebe, who later invented the ECHO classifier, made an early picture division application in the distance differentiating sector. Spatial important Information has been added to the calculations to separate far-flung identifying images, such as region construction, Markovian processes Jackson and Lange], watershed systems, and various evened-out computations. A region is generated using the region creation approach by sharply isolating each neighboring Pixel’s features from the space’s mean. The pixels with tiny differences are distributed throughout the area. As a result, while each zone contains spatially coterminous and homogeneous pixels, there is evident variability at distinct locations [38].

3.3.5 Knowledge-based

In terms of technology, the branch of artificial intelligence has improved in recent years. Rule-based, data-driven, ensemble and reinforcement learning methods are mentioned. The divisions and algorithms for this classification type are listed in this section. In recent years, artificial intelligence has changed the way individuals think about new ideas. Rule-based methodology, data-driven methodologies, ensemble techniques, and reinforcement learning approaches are discussed. This file [39] contains all of the order’s divisions and calculations.

3.3.6 Rule-based methodologies

Rule-based methods, which were the focus of early AI research, extract crucial Information from vast data using master data, agreed-upon rules, and reasoning techniques. The methodology is incredibly logical (predictable from the interaction of rational human thinking) but lacks adaptability, making it analogous to someone born with predetermined knowledge. The criteria cannot be updated once the model has been established, leaving the user powerless to solve new difficulties for which no standards have been developed. The most commonly used principle-based systems for information disclosure from remote detecting data are master frameworks, decision trees, and affiliation rule learning [40].

3.3.7 Distance-based

The least distance classifier (MDC) is an excellent design based on the distance between pixels in the highlight space. It is commonly thought that highlight focuses of the same class are grouped in component space. The mean vector governed by this element focuses as the class’s focal point, and the covariance network represents the scattering of encompassing focuses. Every type has its own set of estimates for focus. The fundamental postulate of the similitude measure is that if the highlight contrasts between two modes are below a given edge, the two modes are believed to be comparable. It addresses a variety of dynamic districts by utilizing a region encompassed by various preparation test centers and calculating test similitude using distance as the significant criterion [37].

3.3.8 Neural based

During the learning process, a neural network structure with numerous layers of nodes (Multilayer Perceptron) sends input observations back and forth until it reaches a termination condition. ANNs were developed as pattern recognition and data analysis tools replicating the brain’s neural storage and analytical functions. Nonparametric ANN approaches, unlike statistical classification methods, do not require prior knowledge of the input data distribution model. ANNs have the advantages of parallel processing, estimating the nonlinear connection between input data and desired outputs, and generalizing quickly. According to multiple earlier research on the classification of multispectral pictures, ANNs outperform standard classification approaches like maximum likelihood classifiers in terms of classification accuracy [41].

3.3.9 Parameter based

Metric approaches such as Support Vector Machines (SVM), Random Forests (RF), and Artificial Neural Networks have all been examined with per-pixel picture order projects (ANN). Well-known grouping strategies have been investigated employing spatial unearthly organization procedures, such as object-based picture inquiry, with significant increases in characterization exactness (OBIA). OBIA-based research, on the other hand, has been limited to the VHR/HR picture files, which are only accessible to specialists [42].

3.3.10 Clustering-based classification

Without human intervention, clustering is an unsupervised machine-learning job that splits data into clusters or groups of related things. It accomplishes this without informing how the groups should appear ahead of time. The technique of putting related elements together is known as “clustering.” This unsupervised machine learning approach looks for commonalities in data points and groups them [43].

3.4 The classifiers of ML for LULC

3.4.1 The supervised classifiers

Algorithms that ‘learn’ patterns in data to predict a discrete class are known as supervised classification approaches. Machine learning techniques are a collection of flexible statistical prediction approaches. The supervised classification use of training data considered representative of each parameter type or unit to be classified is referred to following supervised classifiers:

3.4.1.1 Random Forest –RF

One of the better methods for classification is the RF algorithm. RF is capable of accurately classifying large amounts of data. It is a learning system in which many decision trees are built during training, and the individual trees anticipate the modal outputs. RF is a compilation of Classification and Regression Trees created via discretionary resampling on the readiness set using datasets of equal size to make up a set known as bootstraps. Many bootstraps are used when a tree is built as the test set to avoid joining a specific record from the first dataset. The theory botch as employing a test set of equivalent size as the arrangement set is a measure of the botch speed of the plan of all the test sets. The standard eliminates the need for a different test set. Each tree’s hidden branches vote for one of two classes, and the forest forecasts which class will receive the most votes [44], as described in Figure 6.

Figure 6.

RF system architecture [13].

3.4.1.2 Tree Decision Classification: TDC

A decision tree is an informative model gathered into a decision tree and has center points and constructed edges. The center links two inner issues: leaf center and leaf center points. An inside center addresses a portion of the property, whereas a leaf center addresses a class mark. Figure 7 depicts the planned path from the internal root center to the leaf center, which addresses the request, the fundamental standards, and the gathering measure utilizing decision tree regions as you’d expect from a rule-based classifier. The decision tree is straightforward to comprehend and unravel. It may be combined with various decision techniques to form an outfit learning classifier, such as a self-assertive woods classifier. Observe [39] that the difference between TDC & RF is apparent in Table 4.

Figure 7.

TDC architecture.

	Supervised method	Unsupervised method
Process	In a supervised learning model, input and output variables will be given	In an unsupervised learning model, only input data will be given
Input data	Algorithms are trained using labeled data	Algorithms are used against data that is not labeled
Algorithms used	Support vector machine, neural network, linear and logistics regression, random forest, and classification trees	Unsupervised algorithms can be divided into different categories: cluster algorithms, K-means, hierarchical clustering, etc.
Computational complexity	Supervised learning is a more straightforward method	Unsupervised learning is computationally complex
Use of data	A supervised learning model uses training data to learn a link between input and output	Unsupervised learning does not use output data
Accuracy of results	Highly accurate and trustworthy method	Less accurate and trustworthy method
Real-time learning	The learning method takes place offline	The learning method takes place in real-time
Number of classes	The number of classes is known	The number of classes is not known
Main drawback	Classifying big data can be a real challenge in supervised learning	Cannot get precise information regarding data sorting, and the output as data used in unsupervised learning is labeled and unknown

Table 3.

Comparison of supervised and unsupervised methods.

Decision trees	Random forest
1. When decision trees grow unchecked, they frequently suffer from overfitting.	1. Overfitting is avoided because random forests are generated from subsets of data, and the final output is based on average or majority ratings.
2. A single decision tree is more efficient in terms of computing.	2. In comparison, it is slower.
3. A decision tree will create specific rules for predictions when given a data set with characteristics.	3. Random forest randomly selects data, creates a decision tree, and averages the outputs. It is not based on any formulas.

Table 4.

The difference between TDC & RF.

3.4.1.3 Maximum Likelihood Classifier: MLC

It is one of the most often used remote sensing classification algorithms, in which a pixel with the highest probability is categorized into the appropriate class [40]. It is used in distant detecting order applications. The maximum likelihood classifier MLC computation necessitates proper agent preparation of test information for each category and a detailed assessment of the mean vector and a covariance grid. MLC is a parametric classifier that addresses the inconstancy of courses by applying the covariance grid based on the likelihood that a pixel belongs to a given class. MLC may produce better results than other known parametric orders [37].

3.4.1.4 Spectral Angle Mapper Classification: SAM

It’s a method of comparing photo spectra to a specified range or an automated end member, usually done with a spectrometer in a lab or the field. According to SAM, the data has been reduced to apparent reflectance, according to SAM [41]. Figure 8 shows a technique called Spectral Angle Mapper that maps spectral angles. In an n-D space, where n is the number of bands, this method calculates the spectral angle between both (the unknown and known) spectra as vectors. The size and orientation of each vector are unique. The length of the vector represents the Pixel’s brightness, whereas the vector’s direction represents the Pixel’s spectral characteristic [42].

Figure 8.

Spectral angle mapper (SAM).

3.4.1.5 Support Vector Machine: SVM

SVM is a quantifiable learning Theory-based conditional artistry grouping algorithm. This method is intended to be independent of the dimensionality of the component space [42]. The basic idea behind this arrangement is to use limited pixels to make a choice plane that isolates the classes by enhancing the edge between them. A choice plane separates a group of articles with varying levels of class involvement. The chosen planes may not necessarily be straight lines, as many characterization projects make this impracticable. Tasks that attract separate lines to distinguish various things are known as characterization tasks. Figure 9 shows the situation [42].

Figure 9.

Support Vector Machine (SVM).

3.4.1.6 K-Nearest Neighbor: KNN

KNN was a nonparametric approach utilized in statistical applications. The main idea behind KNN is to discover a collection of k samples in the calibration dataset that are the most comparable to unknown models (based on distance functions, for example) [43]. The response variables (i.e., the class characteristics of the k nearest neighbor) from these k samples are averaged to establish the label (class) of unknown data. As a result, the k plays a critical role in the KNN’s performance for this classifier and is the most vital tuning Parameter for the KNN. A bootstrap technique was used to calculate the parameter k [38].

3.4.1.7 Minimum Distance Classification: MDC

As a regulated arrangement, the spacing between pixels is in the highlight space. It is commonly thought that highlight focuses of the same class are grouped in component space. The mean vector governed by this element focuses as the class’s focal point, and the covariance network represents the scattering of encompassing focuses [39]. Every type has its own set of estimates for focus. The basic premise of the similitude measure is that if the highlight contrasts of the two modes are below a given edge, the models should be comparable [40].

3.4.1.8 Artificial Neural Networks: ANN

ANN Classification is learning to divide data into multiple groups by identifying common characteristics across samples from different classes. ANN of Supervised Learning Classification. Known class labels aid in determining whether or not the system is operating correctly [37]. Information, hidden, and yield layer make up its strategy [41]. The neuron receives the contribution from the left, and each piece of information is multiplied by a weight factor. Learning occurs when the loads in the hub are changed to reduce the gap between the yield hub actuation and the yield [42].

3.4.1.9 Mahalanobis

The Mahalanobis distance is a distance classifier that is sensitive to direction. For each form of input data, it utilizes statistics. While Mahalanobis distance is comparable to maximum likelihood classification, it is quicker since it assumes all class covariance is the same. Because no precise distance cutoff value was applied during software processing, the approach could identify all pixels to the nearest training data [43]. The Mahalanobis distance is a useful multivariate metric for determining between two points. It’s a helpful statistic with applications in multivariate anomaly detection, severely unbalanced dataset classification, and one-class classification. Mahalanobis remote learning has sparked considerable interest, see Figure 10.

Figure 10.

Mahalanobis architecture.

3.4.1.10 Maximum entropy

Maximal Entropy is a group method to the entropy selection criterion that was first proposed. The ensemble classifier’s predictions are used in this method. The collection’s greatest Entropy determines the estimated Uncertainty measure for one instance. (Top) Bounded domain constraint x ∈ [0.7, 1.3] for the traditional equilibrium entropy S eq = ln p eq (x) eq which gives a flat profile and the trajectory entropy S FIT = ρ|∇ 2 |ρ which gives a distribution that scales as p * FIT ∼ cos 2 ((x − μ)π/2 L). (Bottom) Maximum entropy distribution under the constraint on the average x = μ and variance (x − μ) 2 = σ 2 which are equivalent for the static and trajectory information measure as p * FIT = p *, as cleared in Figure 11 [38].

Figure 11.

Maximam entropy.

3.4.1.11 Parallelepiped

A fundamental decision method is used in parallelepiped classification. The decision boundaries in an image data space form an n-dimensional parallelepiped. A standard deviation threshold from the mean for each selected class determines the dimensions of a parallelepiped classifier in Figure 12 [39].

Figure 12.

Parallelepiped.

3.4.1.12 Normal bayes

A Bayesian classifier’s learning module creates a probabilistic model of the characteristics and uses it to predict the classification of a new example. The vector may be used to train a Bayesian classifier. The training data can compute the covariance matrices of the discriminant functions for the abnormal and normal classes [41]. Instead of calculating the maximum of the two discriminant functions, abnormal(x) and standard (x), the choice was made based on the ratio gabnorm(x)/standard (x). The unknown pattern vector is categorized as odd if the ratio is more significant than T; it is expected, as shown in Figure 13 [40].

Figure 13.

Bayesian classifier.

3.4.2 Unsupervised classifiers

Based on geophysical response similarities, unsupervised classification algorithms can objectively classify anomalies into potentially relevant subsurface classifications. Unsupervised classification tries to classify pixels in a remote-sensing image into groups with similar spectral properties without human intervention. Several statistical techniques known as “clustering,” which forms classes of pixels based on their shared spectral signatures, are used to create variety. The following are supervised classifiers [40]:

3.4.2.1 ISOData

It establishes equally distributed class means throughout the data space, then repeatedly clusters the remaining pixels using minimum distance algorithms. Every cycle, the means are recalculated, and the pixels are reclassified. Unsupervised classification with ISODATA calculates class means evenly distributed in the data space, then clusters the remaining pixels using minimum distance approaches. Every iteration recalculates the means and reclassifies pixels based on the new means. The ISODATA algorithm is an iterative method that clusters data components into different classes using Euclidean distance as the similarity measure [37].

3.4.2.2 K-means

It iterates until the best centroid is obtained by calculating the centroids. The data points assigned cluster in the method, resulting in a minimum total squared distance between data points and the centroid. The K-means clustering technique is utilized to locate groupings that have not been explicitly identified in the data [41].

4. Applications for land observation satellites

4.1 LULC change observation

Understanding LULC is essential for managing land resources and evaluating the potential technology [43]. Policymakers can use LULC change detection to understand environmental change dynamics better and assure long-term growth. As a result, LULC feature identification has become a hot topic in research, demanding the creation of a solid and reliable LULC classification method. Land use cover is required to make up the land’s physically existent and apparent surface components [44]. LULC data are necessary for some planning and administration activities, and it is a critical component for illustrating and comprehending the earth as a system. It also plays an essential role in earth-atmosphere interactions [45].

LULC items in any location are essential as a natural and socio-economic component. LULC objects are deep data for various developmental activities on the earth’s surface and their application to human needs in time and space. The land is crucial for humans to carry out any development activities on the planet’s surface, such as agriculture, settlements, and industry. LULC information in the form of maps and statistical data is beneficial for studying land cover patterns, such as agriculture, forestry, economic production, settlements, and environmental studies for spatial planning, management, and land use and exploitation [46]. The LULC classification, without a doubt, plays a critical role in the regional socio-economic development of countries and the management of natural resources, Figure 14.

Figure 14.

Needing to Study LULC.

The LUCC study’s applicability may be used to develop sustainable development in vegetation changes, quantity and quality of water resources, land resources, and coastal management. LULC maps play a crucial and pivotal role in arranging executives and monitoring initiatives. The importance of the discovery of the change of land use land cover LULC based on remote sensing data is the source of information to make appropriate decisions for the benefit of the countries. Disclosure of land change is a factor in conserving land and considering management and development [38]. LULC statistics are required for planning, business, and regulatory needs. The information is also essential for ecological security and spatial arranging because of its spatial nuances. Land use arrangement is indispensable because it gives information that can be utilized to demonstrate, particularly the one managing climate. For example, models manage environmental change and strategies improvements [39].

Land use research and analysis have become prerequisites for proposing a region’s formative activities. In many developing countries, land assets form the foundation for financial development at the national, regional, and local levels. Land usage and land cover data are essential for organizers, decision-makers, and those concerned with land asset management [39]. It enables researchers to look at landscape patterns and features crucial to understanding land size, location, and condition, as well as the ecosystem’s size, structure, and state. Land classification is essential because the usage of a particular piece of land might be connected to considerable price disparities, necessitating a well-defined land categorization. For example, the price (development) of land beneath dwellings may differ significantly from the price (impact) of agricultural land [40]. It provides essential information about human use of the terrain.

Land classification is essential because the usage of a particular piece of land might be connected to considerable price disparities, necessitating a well-defined land categorization. For example, the price (development) of land beneath dwellings may differ significantly from the price (impact) of agricultural land [37], where land use research and analysis have become prerequisites for proposing a region’s formative activities. In many developing countries, land assets form the foundation for financial development at the national, regional, and local levels. Land usage and land cover data are essential for organizers, decision-makers, and those concerned with land asset management [41].

4.1.1 Materials and procedures

4.1.1.1 Sana’a study area

Sana’a is one of the largest cities in Yemen and is in the governorate of the same name as well, and this city is the case study for this article [31]. The city of Sana’a is located at 15°N 44°C or 15.369445 latitudes 44.1191006 with 15°22′ 10.0020’N and 44°11′ 27.6216″E in GPS coordinates [42].

The city of Sana’a Total area is 126 km2 (49 sq. mi), and the population was 2,545,000 issued in 2017. The city has an environment of around 2200 meters above ocean level, see Figure 1. The north-central part of Yemen it’s in a high valley that runs from south to north [12]. With an entire space of 126 Km2 (49 sq. mi), it has a populace of around 3,937,500 (2012). Sana’a’s precipitation is limited to 200 mm/year, while the fading is several times higher.

The average daily sunlight-based irradiance ranges from 800 to 1400 μmol/m², with the month-to-month average air temperature between 22 and 30°C at low humidity levels (35–55%). Its climatic conditions (temperature, sun-based radiation) are ideal for wastewater treatment based on phototrophic [43], as described in Figure 15.

Figure 15.

The location of Yemen Country in the world, Sana’a Governorate, and Sana’a City.

4.1.1.2 Satellite data

This article used the Landsat8 Satellite Sensor (30 m) for LULC mapping & geometrically open-source Landsat8 MSS/TM. The image was obtained from the United States Geological Survey (USGS) of the Sana’a region, a scientific body of the US government. The base map was created from survey photos of the SOI toposheet at a scale of 1:50000 [44]. In this study, the data collected in 1980, 1990,2000,2010,2020 the database details created in Table 5. You can see Sana’a Region on google Maps in Figure 16. The data set of Landsat8 Satellite Sensor (30 m) capture & selection area study with Composite band 432 in Figure 17.

No	period	Satellite	Sensor	Resolution
1	1980	Landsat 5	(TM)	30 m
2	1990	Landsat 5	(TM)	30 m
3	2000	Landsat 7	(ETM+)	30 m
4	2010	Landsat 5	(TM)	30 m
5	2020	Landsat 8	(OLI), (TIRS)	30 m

Table 5.

Database created of images LULCC of Sana’a city.

Figure 16.

Sana’a Region in google map.

Figure 17.

Data set of Landsat8 with Composite band 432.

4.1.1.3 Methodology

The following diagram illustrates the essential steps of this research study in Figure 18.

Figure 18.

Workflow diagram for proposed methodology.

4.1.1.4 Create database

For constructing a database of observation land changes of Sana’a’ in Yemen, the data used in the LULC classification here are 1980,1990,2000,2010,2020 for extracting the differences of decadal period land changes of the region, and the composition of the database is shown in Table 5. Note Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS), Thematic Mapper (TM) Enhanced Thematic Mapper Plus (ETM+), and Multispectral Scanner (MSS).

4.1.1.5 Pre-processing

It is the primary stage and essential task in the process of LULCC, the coordinate reference system for defining and cutting the map into specific areas. The pre-processing procedure identifies the data after it is downloaded from satellites under remote sensing technology. The information subject to pre-processing is divided into the images shown in WGS84 or WGS84 / UTM.

They are pre-processed to contain valid data with a geometrically calibrated reflection often present in the upper atmosphere. The data is not overtly distributed, but its implementation is terminated by organizations responsible for managing satellites. Figure 7 will display the pre-processing corrections for Landsat 8 satellite images in which Band 543 in decadal time 1980,1990,2000,2010,2020.

The images cleared differences in these images map before classification. According to the colors of the Landsat satellite, the region’s red color is vegetation, and the white color is bare land, light gray is land area & network road, and dark gray is built-up area. Initially, the comparison is clear how to land in Sana’a city is changed. The data set of Landsat8 Satellite Sensor (30 m) & selection area study with Composite band 432 cleared in Figure 19.

Figure 19.

Data set of Landsat8, selection area study with composite band 432.

4.1.1.6 Classification for Sana’a city land from 1980 to 2020

As input layers for model processing, there are six samples for six parameters for creating model classes: High Land, Mountains, Land Area, Built-up, Vegetation, and Bare Land. Note to parameters in software SAGA with these models classification in down are seven, but in processing and results in the Parameter are six since merge area vegetation with agriculture land. Create the samples depending on RGB color composites of sentinel-2A images, for example, the class Vegetation (red pixels in color composite RGB = 432), detailed changes in the region. The following details illustrate the critical description of class input in Table 6.

LCLU Class	Description
High Land	High Land Remote may be settlements and clans with a long history and profound loyalties.
Mountains	A mountain is a raised section of the earth’s crust with steep sides and exposed bedrock.
Land Area	The area in square kilometers of the land-based portions of conventional geographic regions is called the land area, which is the population of people. Not contains buildings, maybe streets, parks, roads or buildings crashed down, like this.
Builtup	Built-up areas may be Large buildings, small buildings, settlements, transportation, land, or places containing people like banks, schools, hospitals, etc.
Vegetation	Space containing crops, fields, sparse grassland, a Temperate steppe, and a Temperate meadow.
Bare Land	Bare soil, bare rocks, and land do not contain people like the desert.

Table 6.

Description of LULC classes in the study area.

4.1.1.7 Land changes

Figures 20 and 21 images indicate geomorphological changes in Sana’a in the recent period. After 2010, that change has a role in analyzing this study. This study showed the differences in geomorphology during the mentioned period through the land change classification, which suggests that land use in this region is inappropriate. A database of LULC of Sana’a was created in this work. RF classifier used with Landsat images satellites.

Figure 20.

Buildings & infrastructure of Sana’a city before the conflict [3].

Figure 21.

Post-conflict images show the change of the built-up land to the destroyed land of Sana’a’s city [35].

Such research is necessary for developing nations because it will aid in managing natural resources, where LUCC plays a critical role in regional economic development and natural resource management. Destroyed the country’s infrastructure, preventing Sana’a’s vital economic, social, environmental, health, and agricultural development.

This study shows the detailed Analysis classification for Sana’a city land from 1980 to 2020. LULCC was done in 1980, 1990,2000,2010,2020. It can find LULC classified for Sana’a city, and the categories can be apparent in the differences in land change in Sana’a city as shown in Figure 22.

Figure 22.

Classified map for Sana’a over the five decades (1980–2020).

The summary report compares built-up Areas and Land areas through 1980, 1990,2000,2010,2020. The findings revealed that the political problem began after 2010, as the built-up area decreased on a map in 2010 while the land area increased. I was implying that the poor state of Sana’a city was caused by the war, with increasing built-up area in town resulting in a decrease in land area and the decreasing built-up area in the city’s growing land area. Table 7 and Figure 23 show the opposite situation.

Year	Builtup Area km^2	%	Land Area km^2	%
1980	21,276	9.99%	50,965	48.14%
1990	52,590	24.69%	20,428	19.29%
2000	51,041	23.96%	95,706	9.04%
2010	40,097	18.83%	17,404	16.44%
2020	47,980	22.53%	75,069	7.09%
Total	21,298,644	100.00%	1,058,762	100.00%

Table 7.

Results are calculated for class category land Area & Built-up from 1980 to 2020.

Figure 23.

The chart of results is calculated for class category land Area & Built-up from 1980 to 2020.

The summary report is apparent in Figure 23 to compare built-up Areas and Land areas through 1980, 1990,2000,2010,2020. The findings revealed that the political problem began after 2010, as the built-up area decreased on a map in 2010 while the land area increased. Increasing built-up area in town results in a decrease in land area and the decreasing built-up area in the city growing land area.

4.1.1.8 Results

According to the study’s findings, Table 8 shows the area and percentages of LULC over the decadal period of Sana’a City from 1980 to 2020. All the region’s size factors have been displayed: high land, mountains, land area, built-up area, and vegetation. According to the findings of this study, the built-up area in 1980 was 12.17 percent, and it rose by 34.24 percent in 1990. That is typical, and expansion will continue because of human activity in front of increasing structures and urban development.

No	NAME	1980		1990		2000		2010		2020
		AREA m^2	%	AREA m^2	%	AREA m^2	%	AREA m^2	%	AREA m^2	%
1	High Land	171,612	0.84%	932,976	3.64%	316,881	1.83%	492,039	2.41%	485,370	2.53%
2	Mountains	5,788,260	28.34%	6,728,850	26.23%	5,127,039	29.61%	7,512,750	36.75%	5,990,517	31.24%
3	Land Area	5,096,592	24.95%	2,042,865	7.96%	957,069	5.53%	1,740,402	8.51%	750,699	3.92%
4	Builtup Area	24,867,360	12.17%	8,784,819	34.24%	6,952,491	40.15%	6,325,785	30.94%	8,578,098	44.74%
5	Vegetation	5,334,408	26.11%	5,383,881	20.99%	2,224,908	12.85%	3,333,969	16.31%	2,307,519	12.04%
6	Bare Land	1,549,656	7.59%	1,779,786	6.94%	1,735,965	10.03%	1,037,925	5.08%	1,060,686	5.53%
7	Total of area	20,427,264	100%	25,653,177	100%	1,731,435	100%	20,442,870	100%	19,172,889	100%

Table 8.

Area and percentages LULC for decades period of Sana’a City from 1980 to 2020.

The built-up area was 40.15 percent in 2000, then decreased to 30.94 percent in 2010, which is not typical. The built-up area was 44.74 percent in 2020. Perhaps this is back to political events after 2010 that led the development movement backward in all sectors, including the economy. The remaining analysis parameters had an impact on increasing and decreasing.

The destruction of missiles and the expansion of barren terrain are the main reasons for the shrinking built-up area. The results of land change are mentioned in detail in Figure 24. The area under significant land-use or land-cover classes was calculated for 1980,1990,2000,2010 and 2020. The region’s area in 1980 was 1,867,950,000 km², and in 2020 was 1,497,207,600 Km2. The difference between them is 370,742,400 km², which means the percentage difference is 19.85% of all geographic space of the city. During this period, there has been a persistent reduction in land cover as woodlands expand in cropland and developed regions. Somewhere in the range of 1980 and 2020, of the six significant LULC classes, an extensive increase and decrease have been recorded see Table 8.

Figure 24.

Percentages LULC for decades period of Sana’a City from 1980 to 2020.

The study concluded that human factors and processes have greatly affected the shapes of the earth’s surface in Sana’a by comparing maps for the years 1980 and the year 2020. Human activities have affected the disappearance of many forms of the earth’s surface that contain gains from the Yemeni civilizational heritage, such as castles, forts, and caves, due to human activities and the work of crushers in the mountains. It was reached to create a database for a geomorphological map of the study area. The study recommends valuing biological and human geographical studies to identify the processes and factors affecting the formation of the earth’s surface forms. They benefit from planning and conducting comprehensive development projects and employing them to develop the mountainous heights in Sana’a through building dams and parks and establishing a shelter. The importance of Benefiting from the study of spatial analysis and choosing the optimal site through geographic information systems to make service projects, such as planning to establish a water barrier.

4.1.1.9 Discussion

The study’s findings showed the area and percentages of LULC over the decadal period of Sana’a City from 1980 to 2020. All the region’s size factors have been displayed: high land, mountains, land area, built-up area, and vegetation. According to the findings of this study, the built-up area in 1980 was 12.17%, and it rose by 34.24% in 1990. That is typical, and expansion will continue because of human activity in front of increasing structures and urban development. The built-up area was 40.15% in 2000, then decreased to 30.94% in 2010, which is not typical. The built-up area was 44.74% in 2020. Perhaps this is back to political events after 2010 that led the development movement backward in all sectors, including the economy. The remaining analysis parameters had an impact on increasing and decreasing.

Regarding Sana’a, the region has progressed in urban density, built-up area, and bare ground before 2010 and the opposite after 2010. The area under significant land-use or land-cover classes was calculated for 1980, 1990, 2000, 2010, and 2020. The region’s area in 1980 was 1,867,950,000 m²and in 2020 was 1,497,207,600 m²; the difference between them is 370,742,400 m², which means the percentage difference is 19.85% of all geographic space of the city. During this period, there has been a persistent reduction in land cover as woodlands with attending expansion in cropland and developed region. Somewhere in the range of 1980 and 2020, of the six significant LULC classes, an extensive increase and decrease have been recorded.

This study tried summary of the factors and reasons potentially of the land changes in the Sana’a region is as follows:

1. Events of the war, the asset verification revealed findings that the damage on the ground matches evidence shown in satellite and field photographs. Furthermore, satellite imaging can be used to quickly verify assets when analyzing large-scale damage [24].

2. Human factors for extraction of building materials. Building houses and residential buildings required different building materials, including gravel, which led to the great demand for the development of many mines on those sites and the establishment of crushers that affected the shapes of the Earth’s surface. They worked to drain critical natural resources, which shows human activities and operations in the production areas in the highlands.

3. Erosion and climate factors and Occasional flash floods and potential disasters. Several natural characteristics in several aspects characterize the study area. Climate, one of the biological factors, played a role in the formation and change of these manifestations. Human processes increase as the population grows, reflected in the number of people. Construction, quarrying, and other lands in agriculture and industry are examples of human operations and activities from 1980 to 2020, with an average consumption and conversion of building materials of about five tons.

4.2 Satellite altimetry monitoring changes in mean sea level

The marine gravity field is primarily reliant on satellite altimetry. The accuracy and resolution of the marine gravity field model have been significantly improved due to the development of altimetry missions and advancements in altimeter data processing techniques. However, recovering high-accuracy and high-resolution gravity fields from satellite altimeter data continues to be a difficult task.

Since altimeter data processing techniques are crucial for obtaining precise measurements of sea surface height, these upgraded approaches are then discussed and reviewed with a focus on coastal altimetry. The difficulties in processing altimeter data are also emphasized. The characteristics of gravity recovery methods, including least squares collocation, the inverse Vening Meinesz formula, the inverse Stokes formula, and the inverse Vening Meinesz formula, are also reviewed in the third section. The most recent global marine gravity field models, altimeter data, and processing methods are also shown.

Shipboard gravity measurements also assess the effectiveness of the current global gravity field model. In the low-middle latitude regions, the root means square of the difference between the shipboard gravity from the National Centers for Environmental Information and the global marine gravity model is roughly 5.10 mGal, which is better than the outcome in high-latitude regions. The accuracy of models in the coastal areas still has to be improved, especially within 40 km of the coastline. The SDUST2021GRA model created by the Shandong University of Science and Technology team also showed a fascinating performance. The difficulties in recovering the marine gravity field from satellite altimetry are finally discussed [46].

The sensor aboard an altimetry-focused satellite delivers microwave pulses in the radar frequency range or laser pulses in the optical or infrared spectrum to the ground, which is reflected at the planet’s surface and collects the return signals. The radar observation method is particularly suitable over oceans and open water on land due to the favorable, reflecting characteristics of water. The fundamental measurement is the signal’s round-trip time from the satellite to the water’s surface, which, when multiplied by the speed of light, equals the signal’s round-trip distance. The range, or separation between the satellite and the instantaneous sea surface, is roughly equal to the product of the two-way travel time and the speed of light.

The satellite’s height above a global ellipsoid is calculated from its orbit about a geocentric reference frame (e.g., the International Terrestrial Reference Frame – ITRF).

A laser altimeter, which works similarly to radar technology but employs light pulses, can also determine altitude. Digital elevation models are frequently created using laser altimetry and measuring the elevation change of ice sheets. Their mass balance in response to global warming, satellite radar, and laser altimetry has more recently been used to measure the water level of lakes, rivers, and floodplains on land. Figure 25 indicates Satellite Altimetry Monitoring Changes In Mean Sea Level [48].

Figure 25.

Satellite altimetry monitoring changes in mean sea level [47].

Sea surface height (SSH) readings from satellite altimeters are a regular source of information for tracking ocean processes. It is difficult to completely utilize the available altimeter observations to correctly examine minor mesoscale variations in SSH because, below a wavelength of about 70 km, along-track altimeter measurements frequently exhibit a severe decline in signal-to-noise ratio (SNR).

Although many different strategies have been put forth and used to separate noise from measurements and detect it, no transparent methodology has evolved for systematic use in operational products. The Copernicus Marine Environment Monitoring Service (CMEMS) offers detailed band-pass filtered data to reduce noise contamination of along-track SSH signals to best address this unresolved issue. Users looking to reveal small-scale altimeter signals are thus left to their own devices to devise more creative and appropriate noise-filtering solutions [49].

Here we show that an entirely data-driven strategy is effectively designed and deployed to produce reliable estimates of noise-free sea level anomaly (SLA) signals (Quilfen, 2021). The approach combines a discrete wavelet transform (DWT)-inspired adaptive noise filtering technique with empirical mode decomposition (EMD), which is used to investigate non-stationary and nonlinear processes. It is discovered that this range of mesoscale wavelengths, between 30 and 120 km, better resolves the pattern of SLA variability.

The denoising method, which assumes that the SLA variability is partially the product of a stochastic process, results in a practical uncertainty variable associated with the denoised SLA estimations and considers errors related to the local SNR as well as process uncertainties. The measurements from the missions Jason-3, Sentinel-3, and SARAL/AltiKa are processed and analyzed. Their energy spectrum and seasonal distributions are defined in the small mesoscale domain over the period that is currently accessible. The SASSA (Satellite Altimeter Shortscale Signals Analysis) data set of denoised SLA measurements for three reference altimeter missions has already been shown to yield valuable opportunities to assess global small mesoscale kinetic energy distributions in anticipation of the upcoming SWOT (Surface Water and Ocean Topography) mission data (Quilfen and People, 2021) [47].

5. Challenges of AI techniques for LULC classification

Despite considerable recent progress in AI for LULC, global land-use intensity mapping has faced significant challenges in recent years. Artificial intelligence techniques have spread widely and provided many new solutions to various areas of the natural world and the difficulties of human society. On the other hand, the challenges of artificial intelligence techniques appeared for each field separately.

This section will identify the challenges of artificial intelligence techniques for mapping extensively, with challenges in monitoring how the land cover is classified. As platforms and sensors improve, new issues develop, such as high-dimensional datasets (high spatial resolution and hyperspectral features), sophisticated data structures (nonlinear and overlapping distributions), and the nonlinear optimization problem (high computational complexity) [29].

The complexity of multi-source data exacerbates the difficulty of developing robust and discriminative representations from training data with AI techniques [30]. It might be considered a diverse and significant data processing challenge. Large training samples are necessary for supervised AI systems, generally obtained through time-consuming and labor-intensive processes such as human interpretation of RS products and field surveys. With little training data, developing a robust model of AI-based approaches is a significant difficulty. Techniques for unsupervised AI must be developed.

There are a variety of AI models and frameworks that are both efficient and accurate. Researchers are continually proposing new AI-based change detection systems at the moment. However, it is a significant task to choose an efficient one and ensure its correctness for various applications. In practical applications, AI’s dependability must be considered [31]. Some researchers have looked at these issues and suggested viable solutions, and it will summarize them separately as follows:

5.1 Issue AI’s reliability

When using AI techniques for change detection, factors affecting the reliability of data preparation, model training, change feature extraction, and accuracy evaluation should be considered. The goal is to find the most plausible AI framework for enhancing change detection accuracy. We’ve discussed the issues and promises of AI-based change detection systems in this section and our forecasts for the future [32].

Although many AI-based change detection systems provide the model structure, their trainable parameters are opaque, making it difficult to comprehend why and how they work [33]. AI reliability aims to develop methods for improving the accuracy and interpretability of change detection systems. As a result, it is necessary to build change detection AI that is both resilient and interpretable [34]. Table 9 describes only the approaches that can improve the accuracy of change detection findings from the following areas.

AI’s reliability
1	2	3
Reduce data uncertainty caused by geometric and spectral disparities by eliminating mistakes produced by data sources (such as preprocessing and radiometric correction) or merging different data to improve the original data’s reliability, increasing change detection conclusions’ dependability. A few studies have investigated the influence of registration and algorithm fusion.	Improve AI model interpretability by utilizing a sub-modular model structure, which can help comprehend the overall AI model’s operation principle by understanding the role of each sub-module. R-region-proposals CNN’s component, for example, can be regarded as a generator that predicts object regions.	Improve the durability of AI models by combining many approaches and outcomes.; Ensemble learning is a good strategy for increasing the accuracy of the final output by combining the findings of multiple models.
4	5	6
Reduce noisy points and provide accurate boundaries by including post-processing methods such as the Markov random field, the conditional random field, and level set evolution into the AI model.	To improve the sharpness of change maps, use more suitable detection units. Based on the detection unit of change detection, it may be divided into scene level, patch or super-pixel level, pixel level, and sub-pixel level, from coarse to fine. In terms of dependability, the sub-pixel level is the best option because it avoids the problem of mismatched pixels in RS images. However, it swiftly escalates the level of computational complexity. As a result, the best solution is to employ separate detection units for different land cover types, which requires a well-designed AI model.	To improve the representation of change maps, detect changes in each instance. Change maps include binary, one-class mappings, from–to maps, and instance maps. Although research is still inadequate, the instance change map is more realistic. Because it can provide change information for each instance, it indicates real-world changes. It may also avoid the binary map’s lack of semantic information and the classification system’s restriction of the form–to map, increasing the dependability of the final result.

Table 9.

Directions for solutions issue AI’s reliability.

5.2 Issue AI without supervision

While domain knowledge can aid in constructing representations in classic machine learning methods, AI drives the development of more powerful unsupervised methods. Data can be used to teach unsupervised representation-learning algorithms to learn hierarchical properties. It’s feasible to make data-driven decisions with it [35]. Table 10 summarizes the aspects of unsupervised AI research.

Unsupervised AI
1	2	3
Many researchers have not trained efficient AI models due to a shortage of labeled examples in recent years.; I have put forth a lot of work to solve these issues and have regularly delivered excellent outcomes.; GAN, transfer learning, and other unsupervised AI approaches are developing continually.	Change detection is frequently seen as a low-likelihood issue due to the ambiguity of the change location and direction (i.e., the unaltered change map is significantly more significant than the change). Unsupervised AI algorithms struggle to address this difficulty due to a lack of experience. Although more study is needed to improve performance, weakly- and semi-supervised AI systems are viable alternatives to supervised AI. Nonetheless, a pure unsupervised AI approach to change detection should be the ultimate goal.	One of the motivations for looking into unsupervised AI systems is the lack of training samples or prior knowledge. An excAn excellent alternative strategy is to use crowd-sourced data as a priori knowledge. The Web 2.0 age has arrived on the Internet (emphasizing user-generated content, simplicity of use, participatory culture, and end-user interoperability). For example, OpenStreetMap, a free wiki world map, may provide massive annotation data labeled by volunteers for AI model training. Although the label precision of specific crowd-sourced data is low, the AI model can be trained in a weakly supervised manner to detect changes.

Table 10.

Directions for solutions issue unsupervised AI.

5.3 Issue heterogeneous big information processing

Heterogeneity is a significant property of vast and heterogeneous data, and it complicates the formulation and analysis of change detection findings. SAR, GIS data, high-resolution satellite pictures, and other time and space-measured data are just a few examples of the kind of data that RS technology may provide for change detection. These data of many sorts and formats are difficult to use due to missing values, considerable data redundancy, and unreliability. Furthermore, the generalization ability of current AI systems in RS data processing, particularly heterogeneous significant data processing, must be improved [36]. As a result, we believe the following issues warrant further examination, as in Table 11.

Heterogeneous big data processing
1	2	3
Several AI-based change detection systems based on heterogeneous data have proven effective. However, the number of studies is small. Furthermore, spotting changes across multiple data sources is more important to them than catching data fusion over time. Data fusion theories (i.e., mutual compensation of several types of data) and multi-source data (e.g., optical RS images and DEM) combined with AI techniques can help improve change detection accuracy.	Because current change detection methods rely mainly on detecting 2D data, using 3D data to detect changes in buildings and other structures is also a future research direction. Three-dimensional reconstruction based on oblique images or laser point cloud data and dimensional information integration based on aerial imaging and ground-level street view imagery (i.e., air-ground integration) is two. For recognizing 3D changes, there are currently no sound AI systems.	The application of the AI model is limited since analyzing RS’s extensive data requires many CPU resources. Large-format data processing, for example, is frequently done in blocks, which might lead to edge concerns. A large amount of data needs many trainable parameters in the AI model, resulting in a lengthy training process that uses many computing resources. As a result, the volume of data and the number of trainable parameters must be balanced. They make developing AI-based change detection tools more challenging.

Table 11.

Directions for solutions issue heterogeneous big data processing.

6. Conclusions

Classifiers that create exact LULC maps are in high demand, and dependable Information is required from remotely sensed pictures, even on high-dimensional, complex data. Machine Learning Classifiers have a significant role in giving good classification results. Several aspects influence the accuracy of classified maps, including training sample size, training sample quality, thematic correctness, classifier choice, study region size, etc. Understanding these criteria will aid in achieving the highest classification accuracy feasible for a given need. Big Data challenges arise when classification tasks involving multiple satellite photos and features become computationally intensive.

In recent years, artificial intelligence techniques have spread widely and provided many new solutions to various areas of the natural world and the difficulties of human society. On the other hand, the challenges of artificial intelligence techniques appeared for each field separately. This chapter identified the challenges of artificial intelligence techniques for mapping extensively with challenges in monitoring how the land cover is classified since advances in technologies catalyzed by machine learning and artificial intelligence.

One challenge is the infrastructure, especially the infrastructure of the ancient cities where roadworks were built at different times and with other materials. to face this challenge, it may help with multispectral data technology that can identify more objects and generate more categories. High resolution is the super-spectral for remote sensing data available for urban areas recently. And the accuracy challenge is among the significant challenges in this field. Where noticed was that many researchers used satellite imagery with an accuracy of 30 cm. Also, among the challenges, the challenge in image classification was the weak role of the analyst in the category and the possible classification errors.

The main challenge in LULC changes using remote sensing data to provide accurate and timely geospatial information is clarified as follows. Urban growth has long been considered a sign of regional economic vigor. Still, its benefits increasingly negatively impact the ecosystem and environment, including road traffic, air quality, loss of farming area, social fragmentation, and infrastructure cost. Natural resource management, planning, and monitoring programs depend on accurate information about the land cover in a region. The production of a thematic map from this classification using an image classification is one of the most common applications of remote sensing.

Acknowledgments

Thankful acknowledgments to team IntechOpen for supporting my achievement in this chapter.